Understanding LoRA Rank Trade-offs in Diffusion Model Fine-Tuning
研究通过控制实验分析了不同LoRA秩在扩散模型微调中的性能与成本平衡,结果表明中等秩(如4或8)在保持高质量的同时具有最高的效率。
研究通过控制实验分析了不同LoRA秩在扩散模型微调中的性能与成本平衡,结果表明中等秩(如4或8)在保持高质量的同时具有最高的效率。
本文通过在高斯空间中使用筛估计方法,从有噪声的配对数据中估计最优传输映射,将问题简化为有限维的经验风险最小化。
本文利用图神经网络解决归纳相关聚类问题,通过学习图的结构模式和节点特征,有效处理未见图实例,提高算法的可扩展性和效率。
This study addresses the fragmented understanding of sociotechnical risks in human-AI collaboration, which has hindered the identification of common failure mechanisms and effective interventions. To overcome this limitation, the work proposes a unified lifecycle framework encompassing four phases: task allocation, interaction, feedback, and adoption. Through a cross-domain literature review and conceptual modeling, it synthesizes empirical evidence from healthcare, journalism, education, and scientific research to establish the first comprehensive risk taxonomy. The analysis reveals six core risk clusters—including miscalibrated trust, cognitive overload, and responsibility gaps—and elucidates their cascading interdependencies. By moving beyond isolated risk assessments, this research provides a theoretical foundation for resilient human-AI collaboration and informs end-to-end governance strategies and human-centered AI system design.
Existing knowledge distillation methods for object detection typically treat all instances equally and employ heuristic or teacher-only attention filtering mechanisms, overlooking the student model’s learning dynamics and instance-level variations. This work proposes a learnable instance-aware attention filtering framework that, for the first time, integrates the student model’s dynamic learning state into an instance-level attention mechanism. By introducing a trainable instance selector, the method adaptively reweights the importance of each instance, enabling end-to-end adaptive distillation. This approach departs from conventional static or teacher-dominated filtering paradigms and achieves significant performance gains on KITTI and COCO benchmarks: a GFL ResNet-50 student model improves by 2% mAP without additional computational overhead, outperforming current state-of-the-art methods.
研究通过控制实验分析了不同LoRA秩在扩散模型微调中的性能与成本平衡,结果表明中等秩(如4或8)在保持高质量的同时具有最高的效率。
本文通过在高斯空间中使用筛估计方法,从有噪声的配对数据中估计最优传输映射,将问题简化为有限维的经验风险最小化。
本文利用图神经网络解决归纳相关聚类问题,通过学习图的结构模式和节点特征,有效处理未见图实例,提高算法的可扩展性和效率。
This study addresses the fragmented understanding of sociotechnical risks in human-AI collaboration, which has hindered the identification of common failure mechanisms and effective interventions. To overcome this limitation, the work proposes a unified lifecycle framework encompassing four phases: task allocation, interaction, feedback, and adoption. Through a cross-domain literature review and conceptual modeling, it synthesizes empirical evidence from healthcare, journalism, education, and scientific research to establish the first comprehensive risk taxonomy. The analysis reveals six core risk clusters—including miscalibrated trust, cognitive overload, and responsibility gaps—and elucidates their cascading interdependencies. By moving beyond isolated risk assessments, this research provides a theoretical foundation for resilient human-AI collaboration and informs end-to-end governance strategies and human-centered AI system design.
Existing knowledge distillation methods for object detection typically treat all instances equally and employ heuristic or teacher-only attention filtering mechanisms, overlooking the student model’s learning dynamics and instance-level variations. This work proposes a learnable instance-aware attention filtering framework that, for the first time, integrates the student model’s dynamic learning state into an instance-level attention mechanism. By introducing a trainable instance selector, the method adaptively reweights the importance of each instance, enabling end-to-end adaptive distillation. This approach departs from conventional static or teacher-dominated filtering paradigms and achieves significant performance gains on KITTI and COCO benchmarks: a GFL ResNet-50 student model improves by 2% mAP without additional computational overhead, outperforming current state-of-the-art methods.